Testing 2026 Census messaging

The interventions will be 4 different messages focusing on different behavioural interventions. There will also be a standard control, and an attention control.
Control: A standard ABS message about completing the Census.
Attention Control (AC): A standard ABS message about administrative data.
Intervention 1 (T1): Question and answer format.
Intervention 2 (T1): Case study format.
Intervention 3 (T1): Use of metaphors.
Intervention 4 (T4): Highlighting the benefits.

2025-07-21

2025-07-31

The primary outcome for this trial will be the group mean proportion of correct responses.

We will ask a series of True/False questions of participants. Ten of these questions will ask about Census administrative data, and data and privacy, and 5 will ask general questions about the Census. The order of these questions will be randomised.
Our primary outcome will only use the 10 questions relating to administrative data used by the Census. Responses to each item will be binary (1= correct, 0 = incorrect), and scores are summed (range 0-10). Missing or ‘Don’t know’ responses will be coded as incorrect. At an individual level the raw score will be converted into a proportion (0-1) representing the share of correctly classified items for each person. Individual level scores will be averaged within treatment groups to provide the proportion correct for each arm.

There will be two secondary outcomes, peoples’ willingness to complete the Census and their beliefs in misinformation relating to the Census.

Willingness to complete the Census
At an individual level, we will collect participant’s self-reported willingness to complete the Census. This will be a continuous measure collected by asking participants their willingness on a 11-point scale from 0 (not willing to complete the census) to 10 (willing to complete the Census). Individual level outcomes will be averaged within groups to provide the mean willingness by arm.

Beliefs in Misinformation about the Census
Four of True/False questions asked as part of the primary outcome are based on misinformation about the Census that has been shared on social media. However, we code these responses differently from the primary outcome. Responses to each item will be binary (1= incorrect, 0 = correct), and scores are summed (range 0-4). Missing or ‘Don’t know’ responses will be coded as 0. Like the primary outcome, at an individual level the raw score will be converted into a proportion (0-1) representing the share of correctly classified items for each person. Individual level scores will be averaged within treatment groups to provide the proportion correct for each arm.

This research will consist of a 6-arm randomised controlled trials embedded in a survey delivered online using the Qualtrics platform.

Not available

Randomisation will be conducted in Qualtrics using simple randomisation.

Individual respondents

No

N/A

6000

1000 per arm

This experiment has two families of tests. The first family of tests, test the hypothesis that providing information on data and privacy increases peoples knowledge over the control. The second family of tests, test whether behaviourally informed messages increase knowledge compared with the attention control. Refer the pre-analysis plan for the full set of hypotheses. Given we have two families of tests we have two sets of power calculations. For the first family our minimum effect of interest is 5 percentage points, and for the second family it is 3.5 percentage points. We set α = 0.01, and control for multiple comparisons with a Bonferroni correction. The first family of tests the power is 100%. This represents an overpowered trial for this family of tests. To account for this, we will interpret with caution any effect size below 5 percentage points and report exact p-values. For the second family of tests the power is 96%.